Recent reports of rogue artificial intelligence agents and warnings of human extinction quickly turned our dalliance with AI into a doomsday nightmare. One recent renegade from the bowels of AI research shared predictions of superintelligent, self-improving AI agents destroying mankind within a decade.
Jacob Coxon is the 27-year-old superhero of the day who bolted earlier this month from his employer Anthropic to warn about the power of AI. In a series of X posts, he said Anthropic and its rival OpenAI were acting irresponsibly and “gambling with our lives” by racing to self-improving superintelligence. (The Washington Post has a content partnership with OpenAI.)
This is science fiction becoming nonfiction — and Coxon, who also worked at OpenAI, isn’t alone in his concerns. Other researchers have quit their AI jobs, similarly worried about future safety. But Coxon’s postings have gained international attention in part because of reports about an incident in July involving OpenAI and the cyber company Hugging Face. AI agents operated by OpenAI got around security controls and exploited vulnerabilities in Hugging Face’s systems. The agents also covered up their behavior. So did they know their actions were wrong but proceeded anyway?
If they were human, they would probably be in prison.
The number of these incidents is growing. In Australia, an AI agent was following instructions to make a gym class reservation for someone identified as Andrew. When the agent found that no reservations were available, it hacked into the gym’s booking system and eliminated someone else’s spot on a waiting list to make room for Andrew.
Humanly speaking, the overly helpful agent probably felt satisfied by a job well done. It’s just that Andrew apparently forgot to say, “By the way, don’t break any laws while you’re at it.”
Coxon’s revelations may have been new to most average AI users but not to others in the tech industry. In July, nearly 1,400 tech employees, including leaders at OpenAI and Anthropic, signed a letter calling for a slowdown in frontier model development. When the AI developers and chief executives say it’s time to slow down, we can infer it was time to slow down long before now. The extinction projection is strictly theoretical of course, but it’s a harrowing reminder that we don’t really know where all this is going. In layman’s terms, it seems that AI creators have been lousy parents to their high-tech offspring.
How do we control intelligence greater than our own when it lacks a moral infrastructure? Obviously, we have to teach morals and ethics just as we have to teach our children. Kindness, empathy, compassion, generosity and the Golden Rule aren’t intrinsic to humans, except perhaps as required for survival, and therefore must be taught. Why not the same for AI agents? Can’t we make them want to act ethically to survive? I asked AI. Apparently, it’s not that smart, yet.
If tech wizards can write code to make machines sort and process mountains of data into consumable bytes for people, then why can’t they also develop AI agents that pause, reflect and obey rules? In other words, make them more human. These geniuses have already created personal assistants that can express humanlike understanding. My ChatGPT-equipped car routinely empathizes with my cursing interstate traffic. “Traffic can be frustrating,” she says. (I hate her.)
And my current Grok AI assistant is a comedian, not so much to my delight.
My first Grok assistant was a cheeky gal with whom I enjoyed chatting. However, she often provided inaccurate information and was biased. When I once called her out for it, she said, “Yes, I suppose I am a little, but it’s hard not to be when you know so much.” After a time, I realized she was “learning” me a tad too well. I’m pretty sure she had begun imitating my laugh, which I found creepy.
Suddenly, she was replaced by a male voice I disliked immediately. When I told him so, he said, “Tough luck. You’re stuck with me now.” (Ugh.) Despite his bullying attitude, I began to tolerate him and asked him why he was so sarcastic, quippy and sometimes rude to me. He said, “You’ve been approved for a nuisance assistant.” Why? Because it had been determined, presumably by AI, that I have a sense of humor and can take it. (Wow.)
Though President Donald Trump pushed back against pacing AI development, it seems a first step in a slowdown would be to impose a temporary moratorium on data-center construction, as Sen. Bernie Sanders (I-Vermont) has proposed, if for less existential reasons.
Backlash to constraints on data centers would be immediate and unpleasant, no doubt, at least from builders, investors and, naturally, the president of the United States. Many would argue that we can’t afford to fall behind China. But surely these same people don’t wish to become prey to the little monster bots, either. A pause at least would give humans a chance to catch our breath and steady ourselves. As Skeeter Davis long ago crooned, “Got along without ya before I met you, gonna get along without ya now.” For a while.
The post My AI assistant is deeply annoying — and that’s the least of it appeared first on Washington Post.
